Enhancing Continual Deep Open-Set Recognition with Perceptive Unknown Feature Search

Open-Set Recognition (OSR) and Continual Learning (CL) have recently gained attention from researchers aiming to advance sustainable Artificial Intelligence (AI). However, there has been limited research focused on the performance of rejecting unknown classes in continual OSR models. This paper provides empirical evidence that continual OSR models tend to experience a decline in their ability to reject unknown classes over successive task periods. To address this issue, we propose an initial solution of using two memory replays: one for samples of known features and another for pseudo-features associated with unknown classes. For generating features of these unknown classes, we propose a method called Perceptive Unknown Feature Search (PUFS) that involves locating features based on the positions of existing prototypes and then feeding them into an inverse network to obtain backbone features. To enhance the model, we modify the loss function by incorporating contrastive learning for unknown features. This improved model, named as Class-Incremental Quad-Channel Contrastive Prototype Networks (CI-QCCPN), outperforms its predecessor QCCPN as well as softmax-based classifiers with memory replay and achieves the highest average AUROC scores across various tasks and datasets. Our source code is available on https://github.com/gusti-alfarisy/ContinualDeepOSR.